|
| 1 | +import math |
| 2 | + |
| 3 | +import torch |
| 4 | +import torch.nn.functional as F |
| 5 | + |
| 6 | + |
| 7 | +def _mse_per_sample(pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor: |
| 8 | + """Reproduce the per-sample MSE used in calculate_dpo_loss.""" |
| 9 | + return (pred - target).pow(2).mean(dim=list(range(1, pred.ndim))) |
| 10 | + |
| 11 | + |
| 12 | +def _dpo_loss( |
| 13 | + policy_chosen_logp: torch.Tensor, |
| 14 | + policy_rejected_logp: torch.Tensor, |
| 15 | + ref_chosen_logp: torch.Tensor, |
| 16 | + ref_rejected_logp: torch.Tensor, |
| 17 | + beta: float, |
| 18 | + label_smoothing: float = 0.0, |
| 19 | +) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: |
| 20 | + """Reproduce the DPO loss computation from BaseModelSetup.calculate_dpo_loss.""" |
| 21 | + chosen_ratio = policy_chosen_logp - ref_chosen_logp |
| 22 | + rejected_ratio = policy_rejected_logp - ref_rejected_logp |
| 23 | + logits = beta * (chosen_ratio - rejected_ratio) |
| 24 | + dpo_loss = -F.logsigmoid(logits).mean() |
| 25 | + |
| 26 | + if label_smoothing > 0: |
| 27 | + smooth_loss = -F.logsigmoid(-logits).mean() |
| 28 | + loss = (1 - label_smoothing) * dpo_loss + label_smoothing * smooth_loss |
| 29 | + else: |
| 30 | + loss = dpo_loss |
| 31 | + |
| 32 | + chosen_reward = chosen_ratio.detach().mean() |
| 33 | + rejected_reward = rejected_ratio.detach().mean() |
| 34 | + accuracy = (chosen_ratio > rejected_ratio).float().mean() |
| 35 | + |
| 36 | + return loss, dpo_loss, chosen_reward, rejected_reward, accuracy |
| 37 | + |
| 38 | + |
| 39 | +class TestDPOLossMath: |
| 40 | + def test_beta_zero_gives_log2(self): |
| 41 | + """At beta=0, logits collapse to 0 and -log(sigmoid(0)) = log(2).""" |
| 42 | + B = 4 |
| 43 | + policy_chosen_logp = torch.randn(B) |
| 44 | + policy_rejected_logp = torch.randn(B) |
| 45 | + ref_chosen_logp = torch.randn(B) |
| 46 | + ref_rejected_logp = torch.randn(B) |
| 47 | + |
| 48 | + loss, dpo_loss, _, _, _ = _dpo_loss( |
| 49 | + policy_chosen_logp, policy_rejected_logp, |
| 50 | + ref_chosen_logp, ref_rejected_logp, |
| 51 | + beta=0.0, |
| 52 | + ) |
| 53 | + assert abs(loss.item() - math.log(2)) < 1e-6 |
| 54 | + |
| 55 | + def test_perfect_preference_gives_low_loss(self): |
| 56 | + """When policy strongly prefers chosen over rejected, loss should be low.""" |
| 57 | + B = 4 |
| 58 | + policy_chosen_logp = torch.tensor([0.0] * B) |
| 59 | + policy_rejected_logp = torch.tensor([-10.0] * B) |
| 60 | + ref_chosen_logp = torch.tensor([-5.0] * B) |
| 61 | + ref_rejected_logp = torch.tensor([-5.0] * B) |
| 62 | + |
| 63 | + loss, _, chosen_reward, rejected_reward, accuracy = _dpo_loss( |
| 64 | + policy_chosen_logp, policy_rejected_logp, |
| 65 | + ref_chosen_logp, ref_rejected_logp, |
| 66 | + beta=5000.0, |
| 67 | + ) |
| 68 | + assert accuracy.item() == 1.0 |
| 69 | + assert chosen_reward.item() > rejected_reward.item() |
| 70 | + assert loss.item() < 0.01 |
| 71 | + |
| 72 | + def test_inverted_preference_gives_high_loss(self): |
| 73 | + """When policy prefers rejected over chosen, loss should be high.""" |
| 74 | + B = 4 |
| 75 | + policy_chosen_logp = torch.tensor([-10.0] * B) |
| 76 | + policy_rejected_logp = torch.tensor([0.0] * B) |
| 77 | + ref_chosen_logp = torch.tensor([-5.0] * B) |
| 78 | + ref_rejected_logp = torch.tensor([-5.0] * B) |
| 79 | + |
| 80 | + loss, _, _, _, accuracy = _dpo_loss( |
| 81 | + policy_chosen_logp, policy_rejected_logp, |
| 82 | + ref_chosen_logp, ref_rejected_logp, |
| 83 | + beta=5000.0, |
| 84 | + ) |
| 85 | + assert accuracy.item() == 0.0 |
| 86 | + assert loss.item() > 10.0 |
| 87 | + |
| 88 | + def test_label_smoothing_reduces_extreme_loss(self): |
| 89 | + """Label smoothing should make loss less extreme for both directions.""" |
| 90 | + B = 4 |
| 91 | + policy_chosen_logp = torch.tensor([0.0] * B) |
| 92 | + policy_rejected_logp = torch.tensor([-10.0] * B) |
| 93 | + ref_chosen_logp = torch.tensor([-5.0] * B) |
| 94 | + ref_rejected_logp = torch.tensor([-5.0] * B) |
| 95 | + |
| 96 | + loss_no_smooth, _, _, _, _ = _dpo_loss( |
| 97 | + policy_chosen_logp, policy_rejected_logp, |
| 98 | + ref_chosen_logp, ref_rejected_logp, |
| 99 | + beta=5000.0, label_smoothing=0.0, |
| 100 | + ) |
| 101 | + loss_smooth, _, _, _, _ = _dpo_loss( |
| 102 | + policy_chosen_logp, policy_rejected_logp, |
| 103 | + ref_chosen_logp, ref_rejected_logp, |
| 104 | + beta=5000.0, label_smoothing=0.1, |
| 105 | + ) |
| 106 | + assert loss_smooth.item() > loss_no_smooth.item() |
| 107 | + |
| 108 | + def test_mse_per_sample_reduces_correctly(self): |
| 109 | + """MSE reduction should produce [B] shape from [B, C, H, W].""" |
| 110 | + B, C, H, W = 2, 4, 8, 8 |
| 111 | + pred = torch.randn(B, C, H, W) |
| 112 | + target = torch.randn(B, C, H, W) |
| 113 | + result = _mse_per_sample(pred, target) |
| 114 | + assert result.shape == (B,) |
| 115 | + expected_0 = (pred[0] - target[0]).pow(2).mean() |
| 116 | + assert abs(result[0].item() - expected_0.item()) < 1e-6 |
0 commit comments